DataLoader: ValueError('{} attribute should not be set after {} is 'initialized'

Hi,
I am using a custom loader where after each epoch the training images (stored in RAM) are randomly resampled into patches. For this I reset the samplers of my DataLoader, since this seemed the computationally fastest solution.
In Pytorch 0.4 however, the framework forbids the manual call of setattr with the named ValueError.

My current solution is to comment these lines away and everything works fine, but I wonder why this option is restricted by default? - If there are no known associated problems, would a warning not be more appropriate than an actual error that breaks the pipeline?

Are there better alternatives to obtain a dynamic sampling of training patches from larger image files?

Could you post a small code snippet of your resampling into patches?
I’m not sure, what exactly is done with the sampler.

This part is essentially the resampling, the methods are extensions to a torch.utils.data Dataset.
Since I handle medical data I have to make sure to split test and validation after the different patients - in the code below, named “domain”

The images and ground-truth are stored in self.data

Basically the dataset stores raw images and annotation and I generate a coordinate grid.
The getitem method then samples the positions. Due to the resampling the domains at the different indices might change, so I have to update the Samplers.

def _sample_grid(self):
        grid = []
        domains = []

        def parcellate(img, lbl, domain, index):
            width, height = img.shape[:2]
            stride = self.stride
            size = self.size
            off_x, off_y = np.random.randint(0, stride, size=(2,))
            XX, YY = np.mgrid[off_x:width-size:stride, off_y:height-size:stride]
            doms = domain*np.ones((XX.shape[0]*XX.shape[1],), dtype=np.int32)
            indc = index*np.ones((XX.shape[0]*XX.shape[1],), dtype=np.int32)
            return list(zip(indc, XX.flatten(), YY.flatten())), doms

        for idx, item in enumerate(self.data, 0):
            dom, img, lbl = item
            gri, doms = parcellate(img, lbl, dom, idx)
            assert len(gri) == len(doms)
            grid.extend(gri)
            domains.extend(doms)

        return grid, domains
    
    def resample_grid(self):
        self.grid, self.domains = self._sample_grid()


def __getitem__(self, index):

        ii, xx, yy = self.grid[index]
        dom, img, lbl = self.data[ii]

        img_p = img[xx:xx+self.size, yy:yy+self.size, :]
        lbl_p = lbl[xx:xx+self.size, yy:yy+self.size, :].astype('int32')

        sample = (img_p, lbl_p, ii, self.domains[index])
        if self.transform:
            sample = self.transform(sample)

        return sample

After resampling the Samplers are generated by calling:

def get_datasamplers(dset, slides, batch_size):
    indices = np.arange(len(dset))
    train_idc = []
    valid_idc = []
    test_idc  = []

    for slide_id in np.concatenate(slides):
        if slide_id in slides[0]:
            train_idc.extend(indices[slide_id == dset.domains])
        if slide_id in slides[1]:
            valid_idc.extend(indices[slide_id == dset.domains])
        if slide_id in slides[2]:
            test_idc.extend(indices[slide_id == dset.domains])

    train_smpl = SubsetRandomSampler(train_idc)
    valid_smpl = SubsetRandomSampler(valid_idc)
    test_smpl = SubsetRandomSampler(test_idc)

    train_smpl_batch = BatchSampler(train_smpl, batch_size, False)
    valid_smpl_batch = BatchSampler(valid_smpl, batch_size, False)
    test_smpl_batch = BatchSampler(test_smpl, batch_size, False)

    return train_smpl, train_smpl_batch, valid_smpl, valid_smpl_batch, test_smpl, test_smpl_batch

for me something like

def update_datasamplers(dset, slides, batch_size, batch_samplers: dict):
    indices = np.arange(len(dset))
    train_idc = []
    valid_idc = []
    test_idc  = []

    for slide_id in np.concatenate(slides):
        if slide_id in slides[0]:
            train_idc.extend(indices[slide_id == dset.domains])
        if slide_id in slides[1]:
            valid_idc.extend(indices[slide_id == dset.domains])
        if slide_id in slides[2]:
            test_idc.extend(indices[slide_id == dset.domains])

    batch_samplers["train"].sampler.indices = train_idc
    batch_samplers["valid"].sampler.indices = valid_idc
    batch_samplers["test"].sampler.indices = test_idc

worked since python usually uses something similar to call by reference. Depending on your code (where you want to call update_datasamplers) you could also directly pass the loaders and access the sampler indices via loader.batch_sampler.sampler.indices

This will indeed circumvent the error message, but it feels like this might cause the same effects the framework tries to prevent with the guard in setattr

An alternative is surely to rewrite the part in a custom DataLoader without ‘patching’ around in the framework itself.

However, I would be interested to know, why resetting the Samplers was forbidden in the first place?

Interesting to find this 2-year-old post; I just ran into the same problem. My guess about the reason it’s forbidden is that when you are doing distributed processing, especially with an IterableDataset and/or a DistributedSampler which passes a different part of a dataset to each process, Pytorch can lose track of what has already been assigned if the user is allowed to modify the dataloader’s attributes on the fly.